Hmm, this is a weird one. You may have noticed that mGBA 0.9 introduced a "mid-hblank" phase that deasserts the HBlank field in dispstat at 46 cycles before the end of hblank. This was based in part due to some research I did on hardware and I think a weird reading I found of some really specific thing in GBATEK that I'd have to dig up again. Regardless, it seems I may have inverted the behavior? Do you have info about when HBlank IRQs stop being asserted? Also 46 is a weird number (not divisible by 4) so 48 is probably the accurate number here.

Although the drawing time is only 960 cycles (240*4), the H-Blank flag is "0" for a total of 1006 cycles.

Are you referring to this bit in the DISPSTAT section of GBATEK? I also handle that in NanoBoyAdvance, but slightly differently.

Instead of adding a third "mid-hblank" phase I extended the length of the H-draw phase to 1006 cycles and the H-blank phase is only 226 cycles. I assert H-blank DMA and IRQ after 1006 cycles. VCOUNT is updated at the end of H-blank.

I've made two diagrams to compare mGBA's (to my best understanding) state machine with NBA's :

mGBA

XXXXXXXXXXXXXXXXXXXXXXYYYYYYYYYYZZZZ
                      ^             ^
                      |             |
                      |             |_ Update VCOUNT
                      |
                      |_ Request H-blank DMA and IRQ
Phases:
X = H-draw , H-blank bit = 0 (960 cycles)
Y = H-blank, H-blank bit = 1 (226 cycles)
Z = H-blank flip, H-blank bit = 0 (46 cycles)

NanoBoyAdvance

XXXXXXXXXXXXXXXXXXXXXXXXXXYYYYYYYYYY
                          ^         ^
                          |         |
                          |         |_ Update VCOUNT
                          | 
                          |_ Request H-blank DMA and IRQ

Phases:
X = H-draw , H-blank bit = 0 (1006 cycles)
Y = H-blank, H-blank bit = 1 (226 cycles)

In both cases the H-blank bit is unset for 1006 cycles and set for 226 cycles, just aligned differently to the system clock.
However the major difference is when the H-blank IRQ and DMA are triggered in relation to when VCOUNT is updated.

I have updated my test ROM and uploaded it with source code here:
https://github.com/fleroviux/nba-hw-tests/tree/master/ppu/basic-timing

This is the result mGBA currently gets:

And this is the result on hardware:

From that result it seems like mGBA:

1. asserts H-blank DMA ~46 cycles too early relative to the VCOUNT update (HBL DMA test)
2. as a direct consequence of 1. the V-count coincidence bit is measured to be set and unset ~46 cycles too late relative to the H-blank DMA (STAT VCNT SET and STAT VCNT UNSET)
3. asserts H-blank IRQ ~54?? cycles too early relative to the VCOUNT update (HBL IRQ test)

NanoBoyAdvance "passes" this ROM with a minor few cycles difference in when the H-blank IRQ is asserted, so I'd hope that even though the timing in relation to the system clock might still be incorrect, the events should happen with correct timing in relation to each other.

I don't know what you mean by H-blank IRQs getting unasserted. I don't think it should get unasserted until it has been acknowledged by a write to IF. But I hope that this info and the test ROM are useful.

Yes, I know the numbers are off by two, but for whatever reason your tests didn't pass with the right numbers. It needs closer investigation, but it's much closer now.